A Peek on Numerical Programming in Perl and
Python
E. Christopher Dyken∗
Abstract
In this note we peek at the capabilities for numerical programming in the two
high-level scripting languages Perl and Python, both having numerical libraries
providing increased efficiency for iterating over large arrays of data. We do a
superficial investigation on the efficiency of both languages with and without using
numerical libraries.
Using high-level languages can quite often increase the speed substantially for software
development. High-level scripting languages makes rapid prototyping of new ideas and
concepts possible with a minimal amount of effort. However, one crux of numerical
software is efficient traversal of large amounts of data. High-level languages per se has
a deficiency in the sense that such operations are notoriously slow. To overcome this,
both Perl and Python has add-on libraries providing special data types that can hold
large chunks of data efficient, in regard to both memory usage as well as access speed.
Given one can formulate one’s algorithm as element-by-element operations over
n-dimensional arrays, both Perl and Python provide functionality with performance
comparable to compiled C code. Numerical Python[1] (NumPy) provides fast multidimensional capabilities to Python. A new implementation, numarray[2], is available
as well. Perl has its own counterpart to NumPy, the Perl Data Language[3] (PDL).
PDL brings number-crunching capabilities to Perl as well as an interactive shell and
other goodies. To shed some light on the numerical capabilities of high-level scripting
languages, we have implemented the trapezoidal quadrature rule in Python and Perl,
both with and without add-on libraries, as well as in standard C for reference.
We used a formulation of the quadrature rule which takes advantage of the type of
element-by-element operations over arrays that are optimized by the add-on libraries,
!
Z b
X
f (a) + f (b)
f (x) dx = h
f (a + h · i) −
,
2
a
i=0...N
where h = (b − a)/(N − 1) and N is the number of samples. The idea is to make each
implementation have a code path as similar to each other as possible.
We used three different integrands f1 , f2 and f3 , defined as
∗ Centre
of Mathematics for Applications, University of Oslo
1
f1 (x) = x,
f2 (x) = x2 and
f3 (x) = cos(x2 ) sin(x),
where f1 and f2 are simple polynomials, while f3 is a bit more complex involving
some trigonometric functions.
From this we made the following implementations:
A Standard C implementation was used for reference, compiled with GNU gcc
version 3.2.2 using no optimizations whatsoever. The source code can be found
in Appendix A.1.
A Optimized C implementation, identical to the standard C version, but compiled with -O3 optimization.
A Standard Python implementation was run using Python version 2.3.3. The
source code can be found in Appendix A.2.
A Python with NumPy implementation was run using Python version 2.3.3 and
NumPy version 23.1. The source code can be found in Appendix A.3.
A Python with numarray implementation was run using Python version 2.3.3
and numarray version 0.9. The source code can be found in Appendix A.4.
A Standard Perl implementation was run using Perl version 5.8.0. The source
code can be found in Appendix A.5.
A Perl with PDL implementation was run using Perl version 5.8.0 and PDL
version 2.4.1. The source code can be found in Appendix A.6.
In order to make the tests as fair as possible, the *nix system call gettimeofday
was used since it is available in all languages used in this test. This system call returns
the number of seconds and microseconds of wall-clock time since the epoch. It is questionable to use the wall-clock time and not the processor time spent. However, since
all tests were run on the same system under the same levels of system load (virtually
none) several times, it should give a fair indication.
The experiment was carried out by performing each test 100 times for each combination of sampling density, programming language and integrand on a Intel Pentium
4 running at 3.20 GHz with 2 GB of RAM. The performance of each program is plotted in Figures 1, 2 and 3 where average execution time is plotted versus number of
samples.
From the results, we see that plain vanilla Python and Perl doesn’t present a performance comparable to C. This comes as no surprise. However, by using the extensions, the picture changes. Both numarray and PDL exhibits performance relatively
close to C.
NumPy shows a worse performance on the simplest function, but this gap is reduced when the integrand becomes more complex. Test functions 1 and 2 is composed
2
of only simple floating point arithmetic, and thus most time is spent in the surrounding control structures. However, test function 3 uses cos and sin, which is more
costly and dominates the execution time, which in turn results in a reduction of the gap
between the various implementations.
Thus, from this, we conclude that, given your problem can be formulated mostly
as element-by-element operations on arrays, Python or Perl with add-on libraries could
be a viable alternative to the more traditional approach of C or C++, an alternative that
is definitely worth investigating.
References
[1] The Numerical Python home page,
http://www.numpy.org/
[2] The numarray home page,
http://www.stsci.edu/resources/software_hardware/numarray
[3] The Perl Data Language home page,
http://pdl.perl.org/
3
Test function 1
0.14
C
Python
Perl
Average time (sec)
0.12
0.1
0.08
0.06
0.04
0.02
0
0
10
20
30
40
50
60
70
Thousands of samples
80
90
100
90
100
Test function 1
0.009
C
C -O3
NumPy
numarray
PDL
Average time (sec)
0.008
0.007
0.006
0.005
0.004
0.003
0.002
0.001
0
0
10
20
30
40
50
60
70
Thousands of samples
Figure 1: Numerical integration of f (x) = x.
4
80
Test function 2
0.16
C
Python
Perl
Average time (sec)
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
0
10
20
30
40
50
60
70
Thousands of samples
80
90
100
90
100
Test function 2
0.012
C
C -O3
NumPy
numarray
PDL
Average time (sec)
0.01
0.008
0.006
0.004
0.002
0
0
10
20
30
40
50
60
70
Thousands of samples
Figure 2: Numerical integration of f (x) = x2 .
5
80
Test function 3
Average time (sec)
0.25
C
Python
Perl
0.2
0.15
0.1
0.05
0
0
10
20
30
40
50
60
70
Thousands of samples
80
90
100
90
100
Test function 3
0.04
C
C -O3
NumPy
numarray
PDL
Average time (sec)
0.035
0.03
0.025
0.02
0.015
0.01
0.005
0
0
10
20
30
40
50
60
70
Thousands of samples
80
Figure 3: Numerical integration of f (x) = cos(x2 ) sin(x).
6
A
A.1
Source code
C implementation
#include <stdio.h>
#include <math.h>
#include <sys/time.h>
double integrate(double a, double b, double (*f)(double), int N) {
double h
= (b-a)/((double)N-1.0);
double sum = 0.0;
int i;
for(i=0; i<N; i++)
sum += f(a+h*(double)i);
return h*(sum - 0.5*f(a) - 0.5*f(b));
}
double f1(double t) { return t; }
double f2(double t) { return t*t; }
double f3(double t) { return cos(t*t)*sin(t); }
int main(int argc, char **argv) {
struct timeval begin, end;
double (*f)(double t), res, a=0.0, b=1.0;
int N, M, i;
if(argc != 3) return -1;
switch(atoi(argv[1])) {
case 0: f = f1; break;
case 1: f = f2; break;
case 2: f = f3; break;
}
M = atoi(argv[2]);
for(N=1000; N<100000; N+=1000) {
gettimeofday(&begin, NULL);
for(i=0; i<M; i++)
res = integrate(0,1,f,N);
gettimeofday(&end, NULL);
printf("%f %f\n", res,
((end.tv_sec - begin.tv_sec) +
1.0E-6F*(end.tv_usec - begin.tv_usec))/(double)M);
}
}
7
A.2
Python implementation
import sys;
from math import *;
def integrate(a,b,f,n):
h
= (b-a)/(n-1.0);
sum = 0.0;
for i in range(n):
sum += f(a+(i)*h);
return h*(sum - 0.5*f(a) -0.5*f(b));
if __name__ == ’__main__’:
from time import time;
def f1(x): return x;
def f2(x): return x*x;
def f3(x): return cos(x*x)*sin(x);
if len(sys.argv) == 3:
if int(sys.argv[1]) == 0:
f = f1
elif int(sys.argv[1]) == 1:
f = f2
else:
f = f3
a = 0
b = 1
M = int(sys.argv[2])
for N in range(1000, 100000, 1000):
begin = time()
for i in range(M):
res = integrate(a,b,f,N)
avg = (time()-begin)/M
print "%f %f" % (res, avg)
8
A.3
NumPy implementation
import sys;
from Numeric import *;
def integrate(a,b,f,n):
h
= (b-a)/(n-1.0);
vals = f(h*arange(n));
return h*(sum(vals) - 0.5*vals[0] - 0.5*vals[len(vals)-1]);
if __name__ == ’__main__’:
from time import time;
def f1(x): return x;
def f2(x): return x*x;
def f3(x): return cos(x*x)*sin(x);
if len(sys.argv) == 3:
if int(sys.argv[1]) == 0:
f = f1
elif int(sys.argv[1]) == 1:
f = f2
else:
f = f3
a = 0
b = 1
M = int(sys.argv[2]);
for N in range(1000, 100000, 1000):
begin = time();
for i in range(M):
res = integrate(a,b,f,N);
avg = (time()-begin)/M;
print "%f %f" % (res, avg)
9
A.4
Numarray implementation
import sys;
from numarray import *;
def integrate(a,b,f,n):
h
= (b-a)/(n-1.0);
vals = f(h*arange(n));
return h*(sum(vals) - 0.5*vals[0] - 0.5*vals[len(vals)-1]);
if __name__ == ’__main__’:
from time import time;
def f1(x): return x;
def f2(x): return x*x;
def f3(x): return cos(x*x)*sin(x);
if len(sys.argv) == 3:
if int(sys.argv[1]) == 0:
f = f1
elif int(sys.argv[1]) == 1:
f = f2
else:
f = f3
a = 0
b = 1
M = int(sys.argv[2]);
for N in range(1000, 100000, 1000):
begin = time()
for i in range(M):
res = integrate(a,b,f,N)
avg = (time()-begin)/M
print "%f %f" % (res, avg)
10
A.5
Perl implementation
#!/usr/bin/perl
use strict;
use warnings;
use Time::HiRes qw(gettimeofday tv_interval);
sub integrate {
my ($a, $b, $f_ref, $n) = @_;
my $h = ($b-$a)/($n-1.0);
my $sum = 0;
for my $i (0..$n-1) {
$sum += $f_ref->($a + $h*$i);
}
return $h*( $sum - 0.5*$f_ref->($a) -0.5*$f_ref->($b));
}
die unless @ARGV == 2;
my @F
my
my
my
my
= ( sub {shift; },
sub {my $x=shift; $x*$x; },
sub {my $x=shift; cos($x*$x)*sin($x)} );
$f
= $F[shift];
$M
= shift;
($a, $b) = (0.0, 1.0);
($res, $begin, $end, $avg);
for (my $N=1000; $N<100000; $N+=1000) {
$begin = [gettimeofday()];
for my $i (1..$M) {
$res = integrate($a,$b,$f,$N);
}
$avg = tv_interval($begin)/$M;
print "$res $avg\n";
}
11
A.6
PDL implementation
#!/usr/bin/perl
use strict;
use warnings;
use PDL;
use Time::HiRes qw(gettimeofday tv_interval);
sub integrate {
my ($a, $b, $f_ref, $n) = @_;
my $h = ($b-$a)/($n-1.0);
my $vals = $f_ref->($h*xvals($n));
return $h*(sum($vals) - 0.5*$vals->index(0) - 0.5*$vals->index($n-1));
}
die unless @ARGV == 2;
my @F
my
my
my
my
= ( sub {shift; },
sub {my $x=shift; $x*$x; },
sub {my $x=shift; cos($x*$x)*sin($x)} );
$f
= $F[shift];
$M
= shift;
($a, $b) = (0.0, 1.0);
($res, $begin, $avg);
for(my $N=1000; $N<100000; $N+=1000) {
$begin = [gettimeofday()];
for my $i (1..$M) {
$res = integrate($a,$b,$f,$N);
}
$avg = tv_interval($begin)/$M;
print "$res $avg\n";
}
12